Skip to content

README.md

Latest commit

 

History

History
98 lines (65 loc) · 4.96 KB

README.md

File metadata and controls

98 lines (65 loc) · 4.96 KB
Raw

adhd-style — a SkillWire output style

English · Русский

License: AGPL-3.0 Install Release

ADHD-friendly output styling that actually survives the session: action-first, numbered steps, concrete next actions, time estimates, no preamble. No ADHD diagnosis needed — the style shapes how the agent talks, not what it does.

This repo is the SkillWire catalog of output styles for coding agents — adhd-style is the first entry; more styles will live here under skills/<name>/. The repo name (adhd-opencode) and the skill name (adhd-style) are intentionally independent: npx skills add keys installs and updates by the skill's name, not the repo.

Battle-tested: the ruleset here survived six independent red-team audits against the opencode plugin/skill/gate mechanisms and came out simpler each time.

Why this exists

  • Answer first. The first line is the gist, the command, the path. Someone who reads only it has the answer.
  • Short by default. Say the least that fully answers, then stop. Padding spends attention for nothing.
  • Scannable. One idea per **→** block, bold carrying the whole answer, blank lines between points.
  • Numbers are sacred. State thresholds exactly; never widen a scoped rule; a warning rides with the point it guards.
  • No fake off. Safety exceptions (destructive-action confirm, debug spiral, ambiguity) are always on, even in "concise" mode.

Before / after

The same question — "how do I speed up my TypeScript build?" — without and with the style.

Default adhd-style
A paragraph that opens with context, lists possibilities, buries the recommendation, and closes with "hope this helps" **→ Run npx tsc --noEmit and fix what it names.** then a numbered 1-2-3, a caveat, and one next action.

Measured on the upstream style this derives from: work unchanged (97%/97% hidden tests), output ~43% shorter, the answer lands in the first line 75% of the time vs 3%. (Self-reported by the upstream benchmark; reproduced in attention-span.)

Compatibility

The style body is plain markdown with no harness-specific syntax, so it can be dropped into any agent's rules file. Verified behavior:

Harness Always-on (rules file) On-demand (skill) Notes
opencode ✅ instructions / AGENTS.md ✅ skill tool Re-injected every turn, survives compaction, applies to subagents
Claude Code ✅ CLAUDE.md / output-styles ✅ skill Frontmatter stripped; body plain
Codex / AGENTS.md ✅ append to AGENTS.md — Plain markdown body
Gemini CLI ✅ GEMINI.md — Plain markdown body

Always-on is the reliable path (documented trade-off in INSTALL); on-demand rulesets can drift on long sessions.

Install (opencode)

Always-on (recommended) — the style survives compaction, re-injected every turn

mkdir -p ~/.config/opencode
cp skills/adhd-style/output-style.md ~/.config/opencode/output-style.md

Then add to ~/.config/opencode/opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "instructions": ["~/.config/opencode/output-style.md"]
}

Restart opencode. That's it — every session, every subagent, compaction-proof.

On-demand (skill tool / npx skills)

npx skills add skillwire/adhd-opencode          # latest
npx skills add skillwire/adhd-opencode@v0.1     # pin a release

Then load it in a session with the skill tool, or set a /adhd-style command:

# ~/.config/opencode/commands/adhd-style.md
Use the `adhd-style` skill and apply its ruleset for the rest of this session.

Note: on-demand loading applies for the current session only. On long sessions the ruleset can drift — for a persistent default use the always-on install above.

Contents

File What
skills/adhd-style/SKILL.md The skill (frontmatter + full ruleset) for the skill tool / npx skills
skills/adhd-style/output-style.md The style body, no frontmatter — the single source; drop into instructions for always-on
install/opencode.jsonc.example Minimal wiring example

Tune it

The style is a plain markdown body. Fork, edit skills/adhd-style/output-style.md, keep your copy. Zero dependencies.

License & attribution

AGPL-3.0. This package is a compressed derivative of attention-span's Attention-kind style (AGPL-3.0) folded together with rules from i-have-adhd (MIT). See NOTICE and LICENSE.

For personal use in your own config, nothing further is required. If you redistribute modified versions, AGPL share-alike applies.